Decorrelated-Adversarial-Learning

This is an unofficial pytorch implementation for the 2019 CVPR paper 'Decorrelated Adversarial Learning for Age-Invariant Face Recognition' from Tencent AI Lab'

How to train on your own dataset

If you want to start fresh training, follow the procedures. (or use the backbone and train with your own scritps)

Prepare your data

Folder structure

Since it's under pytorch framework, your dataset should follow ImageFolder requirements. Furthermore, age label for every image is needed. That is, the dataset should look like

 path/to/dataset/train/
                       id0/              
                          000_[age].jpg
                          001_[age].jpg
                          ...
                       id1/              
                          000_[age].jpg
                          001_[age].jpg
                          002_[age].jpg
                          ...

where [age] is the age of that face image. If true age labels are not available, DEX age prediction network can be used to predict apparent ages.

Image sizing

The paper (and this implementation) uses (h, w) = (112, 96). Any face alignment method will do.

Add your dataset metainfo in meta.py

your_dataset_name = {
    'train_root': '/data/your_dataset_name/train',      # root dir for training set
    'val_root': '/data/your_dataset_name/val',          # root dir for validation set (optional, set to None if not available).
    'pat': '_|\.',                                      # pattern for re.split(), which split 000_12.jpg to [000, 12, jpg]
    'pos': 1,                                           # index for `[age]` after split
    'n_cls': 82                                         # number of classes
}

Modify main.py

Open main.py, import your dataset from meta.py, and change dataset to your dataset. Change save to the dir where you want to save your trained weights. By default, weights are saved every epoch.

Off we go

Run the following cmd to start training: python3 main.py ctx epochs /path/to/model_weights where

ctx: device id, ctx < 0 for cpu, ctx-th gpu otherwise.
epochs: how many epochs to train.
/path/to/model_weights: path to trained model weights if you want to resume training.

Performance

I trained on a dataset composed by many public cross-age datasets and VGG-face2 general dataset. Best verification accuracy on CSCD_VS was 98.8%.

Keeps updating...